Merge pull request #1081 from router-for-me/thinking

Refine thinking validation and cross‑provider payload conversion

internal/logging/global_logger.go

@@ -30,7 +30,7 @@ var (
 type LogFormatter struct{}
 
 // logFieldOrder defines the display order for common log fields.
-var logFieldOrder = []string{"provider", "model", "mode", "budget", "level", "original_value", "min", "max", "clamped_to", "error"}
+var logFieldOrder = []string{"provider", "model", "mode", "budget", "level", "original_value", "original_level", "min", "max", "clamped_to", "error"}
 
 // Format renders a single log entry with custom formatting.
 func (m *LogFormatter) Format(entry *log.Entry) ([]byte, error) {

internal/runtime/executor/aistudio_executor.go

@@ -393,7 +393,7 @@ func (e *AIStudioExecutor) translateRequest(req cliproxyexecutor.Request, opts c
 	}
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, stream)
 	payload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), stream)
-	payload, err := thinking.ApplyThinking(payload, req.Model, "gemini")
+	payload, err := thinking.ApplyThinking(payload, req.Model, from.String(), to.String())
 	if err != nil {
 		return nil, translatedPayload{}, err
 	}

internal/runtime/executor/antigravity_executor.go

@@ -137,7 +137,7 @@ func (e *AntigravityExecutor) Execute(ctx context.Context, auth *cliproxyauth.Au
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
 	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	translated, err = thinking.ApplyThinking(translated, req.Model, "antigravity")
+	translated, err = thinking.ApplyThinking(translated, req.Model, from.String(), to.String())
 	if err != nil {
 		return resp, err
 	}
@@ -256,7 +256,7 @@ func (e *AntigravityExecutor) executeClaudeNonStream(ctx context.Context, auth *
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
 	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 
-	translated, err = thinking.ApplyThinking(translated, req.Model, "antigravity")
+	translated, err = thinking.ApplyThinking(translated, req.Model, from.String(), to.String())
 	if err != nil {
 		return resp, err
 	}
@@ -622,7 +622,7 @@ func (e *AntigravityExecutor) ExecuteStream(ctx context.Context, auth *cliproxya
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
 	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 
-	translated, err = thinking.ApplyThinking(translated, req.Model, "antigravity")
+	translated, err = thinking.ApplyThinking(translated, req.Model, from.String(), to.String())
 	if err != nil {
 		return nil, err
 	}
@@ -802,7 +802,7 @@ func (e *AntigravityExecutor) CountTokens(ctx context.Context, auth *cliproxyaut
 	// Prepare payload once (doesn't depend on baseURL)
 	payload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	payload, err := thinking.ApplyThinking(payload, req.Model, "antigravity")
+	payload, err := thinking.ApplyThinking(payload, req.Model, from.String(), to.String())
 	if err != nil {
 		return cliproxyexecutor.Response{}, err
 	}

internal/runtime/executor/claude_executor.go

@@ -106,7 +106,7 @@ func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), stream)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 
-	body, err = thinking.ApplyThinking(body, req.Model, "claude")
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String())
 	if err != nil {
 		return resp, err
 	}
@@ -239,7 +239,7 @@ func (e *ClaudeExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 
-	body, err = thinking.ApplyThinking(body, req.Model, "claude")
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String())
 	if err != nil {
 		return nil, err
 	}

internal/runtime/executor/codex_executor.go

@@ -96,7 +96,7 @@ func (e *CodexExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 	body = sdktranslator.TranslateRequest(from, to, baseModel, body, false)
 	body = misc.StripCodexUserAgent(body)
 
-	body, err = thinking.ApplyThinking(body, req.Model, "codex")
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String())
 	if err != nil {
 		return resp, err
 	}
@@ -208,7 +208,7 @@ func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 	body = sdktranslator.TranslateRequest(from, to, baseModel, body, true)
 	body = misc.StripCodexUserAgent(body)
 
-	body, err = thinking.ApplyThinking(body, req.Model, "codex")
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String())
 	if err != nil {
 		return nil, err
 	}
@@ -316,7 +316,7 @@ func (e *CodexExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth
 	body = sdktranslator.TranslateRequest(from, to, baseModel, body, false)
 	body = misc.StripCodexUserAgent(body)
 
-	body, err := thinking.ApplyThinking(body, req.Model, "codex")
+	body, err := thinking.ApplyThinking(body, req.Model, from.String(), to.String())
 	if err != nil {
 		return cliproxyexecutor.Response{}, err
 	}

internal/runtime/executor/gemini_cli_executor.go

@@ -123,7 +123,7 @@ func (e *GeminiCLIExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
 	basePayload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	basePayload, err = thinking.ApplyThinking(basePayload, req.Model, "gemini-cli")
+	basePayload, err = thinking.ApplyThinking(basePayload, req.Model, from.String(), to.String())
 	if err != nil {
 		return resp, err
 	}
@@ -272,7 +272,7 @@ func (e *GeminiCLIExecutor) ExecuteStream(ctx context.Context, auth *cliproxyaut
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
 	basePayload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 
-	basePayload, err = thinking.ApplyThinking(basePayload, req.Model, "gemini-cli")
+	basePayload, err = thinking.ApplyThinking(basePayload, req.Model, from.String(), to.String())
 	if err != nil {
 		return nil, err
 	}
@@ -479,7 +479,7 @@ func (e *GeminiCLIExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.
 	for range models {
 		payload := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-		payload, err = thinking.ApplyThinking(payload, req.Model, "gemini-cli")
+		payload, err = thinking.ApplyThinking(payload, req.Model, from.String(), to.String())
 		if err != nil {
 			return cliproxyexecutor.Response{}, err
 		}

internal/runtime/executor/gemini_executor.go

@@ -120,7 +120,7 @@ func (e *GeminiExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, r
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	body, err = thinking.ApplyThinking(body, req.Model, "gemini")
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String())
 	if err != nil {
 		return resp, err
 	}
@@ -222,7 +222,7 @@ func (e *GeminiExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.A
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 
-	body, err = thinking.ApplyThinking(body, req.Model, "gemini")
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String())
 	if err != nil {
 		return nil, err
 	}
@@ -338,7 +338,7 @@ func (e *GeminiExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Aut
 	to := sdktranslator.FromString("gemini")
 	translatedReq := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	translatedReq, err := thinking.ApplyThinking(translatedReq, req.Model, "gemini")
+	translatedReq, err := thinking.ApplyThinking(translatedReq, req.Model, from.String(), to.String())
 	if err != nil {
 		return cliproxyexecutor.Response{}, err
 	}

internal/runtime/executor/gemini_vertex_executor.go

@@ -170,7 +170,7 @@ func (e *GeminiVertexExecutor) executeWithServiceAccount(ctx context.Context, au
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	body, err = thinking.ApplyThinking(body, req.Model, "gemini")
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String())
 	if err != nil {
 		return resp, err
 	}
@@ -272,7 +272,7 @@ func (e *GeminiVertexExecutor) executeWithAPIKey(ctx context.Context, auth *clip
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, false)
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	body, err = thinking.ApplyThinking(body, req.Model, "gemini")
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String())
 	if err != nil {
 		return resp, err
 	}
@@ -375,7 +375,7 @@ func (e *GeminiVertexExecutor) executeStreamWithServiceAccount(ctx context.Conte
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 
-	body, err = thinking.ApplyThinking(body, req.Model, "gemini")
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String())
 	if err != nil {
 		return nil, err
 	}
@@ -494,7 +494,7 @@ func (e *GeminiVertexExecutor) executeStreamWithAPIKey(ctx context.Context, auth
 	originalTranslated := sdktranslator.TranslateRequest(from, to, baseModel, originalPayload, true)
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 
-	body, err = thinking.ApplyThinking(body, req.Model, "gemini")
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String())
 	if err != nil {
 		return nil, err
 	}
@@ -605,7 +605,7 @@ func (e *GeminiVertexExecutor) countTokensWithServiceAccount(ctx context.Context
 
 	translatedReq := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	translatedReq, err := thinking.ApplyThinking(translatedReq, req.Model, "gemini")
+	translatedReq, err := thinking.ApplyThinking(translatedReq, req.Model, from.String(), to.String())
 	if err != nil {
 		return cliproxyexecutor.Response{}, err
 	}
@@ -689,7 +689,7 @@ func (e *GeminiVertexExecutor) countTokensWithAPIKey(ctx context.Context, auth *
 
 	translatedReq := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 
-	translatedReq, err := thinking.ApplyThinking(translatedReq, req.Model, "gemini")
+	translatedReq, err := thinking.ApplyThinking(translatedReq, req.Model, from.String(), to.String())
 	if err != nil {
 		return cliproxyexecutor.Response{}, err
 	}

internal/runtime/executor/iflow_executor.go

@@ -92,7 +92,7 @@ func (e *IFlowExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, re
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 
-	body, err = thinking.ApplyThinking(body, req.Model, "iflow")
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), "iflow")
 	if err != nil {
 		return resp, err
 	}
@@ -190,7 +190,7 @@ func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 
-	body, err = thinking.ApplyThinking(body, req.Model, "iflow")
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), "iflow")
 	if err != nil {
 		return nil, err
 	}

internal/runtime/executor/openai_compat_executor.go

@@ -92,7 +92,7 @@ func (e *OpenAICompatExecutor) Execute(ctx context.Context, auth *cliproxyauth.A
 	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), opts.Stream)
 	translated = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", translated, originalTranslated)
 
-	translated, err = thinking.ApplyThinking(translated, req.Model, "openai")
+	translated, err = thinking.ApplyThinking(translated, req.Model, from.String(), to.String())
 	if err != nil {
 		return resp, err
 	}
@@ -187,7 +187,7 @@ func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxy
 	translated := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 	translated = applyPayloadConfigWithRoot(e.cfg, baseModel, to.String(), "", translated, originalTranslated)
 
-	translated, err = thinking.ApplyThinking(translated, req.Model, "openai")
+	translated, err = thinking.ApplyThinking(translated, req.Model, from.String(), to.String())
 	if err != nil {
 		return nil, err
 	}
@@ -297,7 +297,7 @@ func (e *OpenAICompatExecutor) CountTokens(ctx context.Context, auth *cliproxyau
 
 	modelForCounting := baseModel
 
-	translated, err := thinking.ApplyThinking(translated, req.Model, "openai")
+	translated, err := thinking.ApplyThinking(translated, req.Model, from.String(), to.String())
 	if err != nil {
 		return cliproxyexecutor.Response{}, err
 	}

internal/runtime/executor/qwen_executor.go

@@ -86,7 +86,7 @@ func (e *QwenExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), false)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 
-	body, err = thinking.ApplyThinking(body, req.Model, "openai")
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String())
 	if err != nil {
 		return resp, err
 	}
@@ -172,7 +172,7 @@ func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 	body := sdktranslator.TranslateRequest(from, to, baseModel, bytes.Clone(req.Payload), true)
 	body, _ = sjson.SetBytes(body, "model", baseModel)
 
-	body, err = thinking.ApplyThinking(body, req.Model, "openai")
+	body, err = thinking.ApplyThinking(body, req.Model, from.String(), to.String())
 	if err != nil {
 		return nil, err
 	}

internal/thinking/apply.go

@@ -2,6 +2,8 @@
 package thinking
 
 import (
+	"strings"
+
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	log "github.com/sirupsen/logrus"
 	"github.com/tidwall/gjson"
@@ -59,7 +61,8 @@ func IsUserDefinedModel(modelInfo *registry.ModelInfo) bool {
 // Parameters:
 //   - body: Original request body JSON
 //   - model: Model name, optionally with thinking suffix (e.g., "claude-sonnet-4-5(16384)")
-//   - provider: Provider name (gemini, gemini-cli, antigravity, claude, openai, codex, iflow)
+//   - fromFormat: Source request format (e.g., openai, codex, gemini)
+//   - toFormat: Target provider format for the request body (gemini, gemini-cli, antigravity, claude, openai, codex, iflow)
 //
 // Returns:
 //   - Modified request body JSON with thinking configuration applied
@@ -76,16 +79,21 @@ func IsUserDefinedModel(modelInfo *registry.ModelInfo) bool {
 // Example:
 //
 //	// With suffix - suffix config takes priority
-//	result, err := thinking.ApplyThinking(body, "gemini-2.5-pro(8192)", "gemini")
+//	result, err := thinking.ApplyThinking(body, "gemini-2.5-pro(8192)", "gemini", "gemini")
 //
 //	// Without suffix - uses body config
-//	result, err := thinking.ApplyThinking(body, "gemini-2.5-pro", "gemini")
+//	result, err := thinking.ApplyThinking(body, "gemini-2.5-pro", "gemini", "gemini")
-func ApplyThinking(body []byte, model string, provider string) ([]byte, error) {
+func ApplyThinking(body []byte, model string, fromFormat string, toFormat string) ([]byte, error) {
+	providerFormat := strings.ToLower(strings.TrimSpace(toFormat))
+	fromFormat = strings.ToLower(strings.TrimSpace(fromFormat))
+	if fromFormat == "" {
+		fromFormat = providerFormat
+	}
 	// 1. Route check: Get provider applier
-	applier := GetProviderApplier(provider)
+	applier := GetProviderApplier(providerFormat)
 	if applier == nil {
 		log.WithFields(log.Fields{
-			"provider": provider,
+			"provider": providerFormat,
 			"model":    model,
 		}).Debug("thinking: unknown provider, passthrough |")
 		return body, nil
@@ -100,19 +108,19 @@ func ApplyThinking(body []byte, model string, provider string) ([]byte, error) {
 	// Unknown models are treated as user-defined so thinking config can still be applied.
 	// The upstream service is responsible for validating the configuration.
 	if IsUserDefinedModel(modelInfo) {
-		return applyUserDefinedModel(body, modelInfo, provider, suffixResult)
+		return applyUserDefinedModel(body, modelInfo, fromFormat, providerFormat, suffixResult)
 	}
 	if modelInfo.Thinking == nil {
-		config := extractThinkingConfig(body, provider)
+		config := extractThinkingConfig(body, providerFormat)
 		if hasThinkingConfig(config) {
 			log.WithFields(log.Fields{
 				"model":    baseModel,
-				"provider": provider,
+				"provider": providerFormat,
 			}).Debug("thinking: model does not support thinking, stripping config |")
-			return StripThinkingConfig(body, provider), nil
+			return StripThinkingConfig(body, providerFormat), nil
 		}
 		log.WithFields(log.Fields{
-			"provider": provider,
+			"provider": providerFormat,
 			"model":    baseModel,
 		}).Debug("thinking: model does not support thinking, passthrough |")
 		return body, nil
@@ -121,19 +129,19 @@ func ApplyThinking(body []byte, model string, provider string) ([]byte, error) {
 	// 4. Get config: suffix priority over body
 	var config ThinkingConfig
 	if suffixResult.HasSuffix {
-		config = parseSuffixToConfig(suffixResult.RawSuffix, provider, model)
+		config = parseSuffixToConfig(suffixResult.RawSuffix, providerFormat, model)
 		log.WithFields(log.Fields{
-			"provider": provider,
+			"provider": providerFormat,
 			"model":    model,
 			"mode":     config.Mode,
 			"budget":   config.Budget,
 			"level":    config.Level,
 		}).Debug("thinking: config from model suffix |")
 	} else {
-		config = extractThinkingConfig(body, provider)
+		config = extractThinkingConfig(body, providerFormat)
 		if hasThinkingConfig(config) {
 			log.WithFields(log.Fields{
-				"provider": provider,
+				"provider": providerFormat,
 				"model":    modelInfo.ID,
 				"mode":     config.Mode,
 				"budget":   config.Budget,
@@ -144,17 +152,17 @@ func ApplyThinking(body []byte, model string, provider string) ([]byte, error) {
 
 	if !hasThinkingConfig(config) {
 		log.WithFields(log.Fields{
-			"provider": provider,
+			"provider": providerFormat,
 			"model":    modelInfo.ID,
 		}).Debug("thinking: no config found, passthrough |")
 		return body, nil
 	}
 
 	// 5. Validate and normalize configuration
-	validated, err := ValidateConfig(config, modelInfo, provider)
+	validated, err := ValidateConfig(config, modelInfo, fromFormat, providerFormat)
 	if err != nil {
 		log.WithFields(log.Fields{
-			"provider": provider,
+			"provider": providerFormat,
 			"model":    modelInfo.ID,
 			"error":    err.Error(),
 		}).Warn("thinking: validation failed |")
@@ -167,14 +175,14 @@ func ApplyThinking(body []byte, model string, provider string) ([]byte, error) {
 	// Defensive check: ValidateConfig should never return (nil, nil)
 	if validated == nil {
 		log.WithFields(log.Fields{
-			"provider": provider,
+			"provider": providerFormat,
 			"model":    modelInfo.ID,
 		}).Warn("thinking: ValidateConfig returned nil config without error, passthrough |")
 		return body, nil
 	}
 
 	log.WithFields(log.Fields{
-		"provider": provider,
+		"provider": providerFormat,
 		"model":    modelInfo.ID,
 		"mode":     validated.Mode,
 		"budget":   validated.Budget,
@@ -228,7 +236,7 @@ func parseSuffixToConfig(rawSuffix, provider, model string) ThinkingConfig {
 
 // applyUserDefinedModel applies thinking configuration for user-defined models
 // without ThinkingSupport validation.
-func applyUserDefinedModel(body []byte, modelInfo *registry.ModelInfo, provider string, suffixResult SuffixResult) ([]byte, error) {
+func applyUserDefinedModel(body []byte, modelInfo *registry.ModelInfo, fromFormat, toFormat string, suffixResult SuffixResult) ([]byte, error) {
 	// Get model ID for logging
 	modelID := ""
 	if modelInfo != nil {
@@ -240,39 +248,57 @@ func applyUserDefinedModel(body []byte, modelInfo *registry.ModelInfo, provider
 	// Get config: suffix priority over body
 	var config ThinkingConfig
 	if suffixResult.HasSuffix {
-		config = parseSuffixToConfig(suffixResult.RawSuffix, provider, modelID)
+		config = parseSuffixToConfig(suffixResult.RawSuffix, toFormat, modelID)
 	} else {
-		config = extractThinkingConfig(body, provider)
+		config = extractThinkingConfig(body, toFormat)
 	}
 
 	if !hasThinkingConfig(config) {
 		log.WithFields(log.Fields{
 			"model":    modelID,
-			"provider": provider,
+			"provider": toFormat,
 		}).Debug("thinking: user-defined model, passthrough (no config) |")
 		return body, nil
 	}
 
-	applier := GetProviderApplier(provider)
+	applier := GetProviderApplier(toFormat)
 	if applier == nil {
 		log.WithFields(log.Fields{
 			"model":    modelID,
-			"provider": provider,
+			"provider": toFormat,
 		}).Debug("thinking: user-defined model, passthrough (unknown provider) |")
 		return body, nil
 	}
 
 	log.WithFields(log.Fields{
-		"provider": provider,
+		"provider": toFormat,
 		"model":    modelID,
 		"mode":     config.Mode,
 		"budget":   config.Budget,
 		"level":    config.Level,
 	}).Debug("thinking: applying config for user-defined model (skip validation)")
 
+	config = normalizeUserDefinedConfig(config, fromFormat, toFormat)
 	return applier.Apply(body, config, modelInfo)
 }
 
+func normalizeUserDefinedConfig(config ThinkingConfig, fromFormat, toFormat string) ThinkingConfig {
+	if config.Mode != ModeLevel {
+		return config
+	}
+	if !isBudgetBasedProvider(toFormat) || !isLevelBasedProvider(fromFormat) {
+		return config
+	}
+	budget, ok := ConvertLevelToBudget(string(config.Level))
+	if !ok {
+		return config
+	}
+	config.Mode = ModeBudget
+	config.Budget = budget
+	config.Level = ""
+	return config
+}
+
 // extractThinkingConfig extracts provider-specific thinking config from request body.
 func extractThinkingConfig(body []byte, provider string) ThinkingConfig {
 	if len(body) == 0 || !gjson.ValidBytes(body) {
@@ -289,7 +315,11 @@ func extractThinkingConfig(body []byte, provider string) ThinkingConfig {
 	case "codex":
 		return extractCodexConfig(body)
 	case "iflow":
-		return extractIFlowConfig(body)
+		config := extractIFlowConfig(body)
+		if hasThinkingConfig(config) {
+			return config
+		}
+		return extractOpenAIConfig(body)
 	default:
 		return ThinkingConfig{}
 	}

internal/thinking/errors.go

@@ -24,6 +24,10 @@ const (
 	// Example: using level with a budget-only model
 	ErrLevelNotSupported ErrorCode = "LEVEL_NOT_SUPPORTED"
 
+	// ErrBudgetOutOfRange indicates the budget value is outside model range.
+	// Example: budget 64000 exceeds max 20000
+	ErrBudgetOutOfRange ErrorCode = "BUDGET_OUT_OF_RANGE"
+
 	// ErrProviderMismatch indicates the provider does not match the model.
 	// Example: applying Claude format to a Gemini model
 	ErrProviderMismatch ErrorCode = "PROVIDER_MISMATCH"

internal/thinking/strip.go

@@ -27,28 +27,32 @@ func StripThinkingConfig(body []byte, provider string) []byte {
 		return body
 	}
 
+	var paths []string
 	switch provider {
 	case "claude":
-		result, _ := sjson.DeleteBytes(body, "thinking")
+		paths = []string{"thinking"}
-		return result
 	case "gemini":
-		result, _ := sjson.DeleteBytes(body, "generationConfig.thinkingConfig")
+		paths = []string{"generationConfig.thinkingConfig"}
-		return result
 	case "gemini-cli", "antigravity":
-		result, _ := sjson.DeleteBytes(body, "request.generationConfig.thinkingConfig")
+		paths = []string{"request.generationConfig.thinkingConfig"}
-		return result
 	case "openai":
-		result, _ := sjson.DeleteBytes(body, "reasoning_effort")
+		paths = []string{"reasoning_effort"}
-		return result
 	case "codex":
-		result, _ := sjson.DeleteBytes(body, "reasoning.effort")
+		paths = []string{"reasoning.effort"}
-		return result
 	case "iflow":
-		result, _ := sjson.DeleteBytes(body, "chat_template_kwargs.enable_thinking")
+		paths = []string{
-		result, _ = sjson.DeleteBytes(result, "chat_template_kwargs.clear_thinking")
+			"chat_template_kwargs.enable_thinking",
-		result, _ = sjson.DeleteBytes(result, "reasoning_split")
+			"chat_template_kwargs.clear_thinking",
-		return result
+			"reasoning_split",
+			"reasoning_effort",
+		}
 	default:
 		return body
 	}
+
+	result := body
+	for _, path := range paths {
+		result, _ = sjson.DeleteBytes(result, path)
+	}
+	return result
 }

internal/thinking/validate.go

@@ -9,64 +9,6 @@ import (
 	log "github.com/sirupsen/logrus"
 )
 
-// ClampBudget clamps a budget value to the model's supported range.
-//
-// Logging:
-//   - Warn when value=0 but ZeroAllowed=false
-//   - Debug when value is clamped to min/max
-//
-// Fields: provider, model, original_value, clamped_to, min, max
-func ClampBudget(value int, modelInfo *registry.ModelInfo, provider string) int {
-	model := "unknown"
-	support := (*registry.ThinkingSupport)(nil)
-	if modelInfo != nil {
-		if modelInfo.ID != "" {
-			model = modelInfo.ID
-		}
-		support = modelInfo.Thinking
-	}
-	if support == nil {
-		return value
-	}
-
-	// Auto value (-1) passes through without clamping.
-	if value == -1 {
-		return value
-	}
-
-	min := support.Min
-	max := support.Max
-	if value == 0 && !support.ZeroAllowed {
-		log.WithFields(log.Fields{
-			"provider":       provider,
-			"model":          model,
-			"original_value": value,
-			"clamped_to":     min,
-			"min":            min,
-			"max":            max,
-		}).Warn("thinking: budget zero not allowed |")
-		return min
-	}
-
-	// Some models are level-only and do not define numeric budget ranges.
-	if min == 0 && max == 0 {
-		return value
-	}
-
-	if value < min {
-		if value == 0 && support.ZeroAllowed {
-			return 0
-		}
-		logClamp(provider, model, value, min, min, max)
-		return min
-	}
-	if value > max {
-		logClamp(provider, model, value, max, min, max)
-		return max
-	}
-	return value
-}
-
 // ValidateConfig validates a thinking configuration against model capabilities.
 //
 // This function performs comprehensive validation:
@@ -74,10 +16,14 @@ func ClampBudget(value int, modelInfo *registry.ModelInfo, provider string) int
 //   - Auto-converts between Budget and Level formats based on model capability
 //   - Validates that requested level is in the model's supported levels list
 //   - Clamps budget values to model's allowed range
+//   - When converting Budget -> Level for level-only models, clamps the derived standard level to the nearest supported level
+//     (special values none/auto are preserved)
 //
 // Parameters:
 //   - config: The thinking configuration to validate
 //   - support: Model's ThinkingSupport properties (nil means no thinking support)
+//   - fromFormat: Source provider format (used to determine strict validation rules)
+//   - toFormat: Target provider format
 //
 // Returns:
 //   - Normalized ThinkingConfig with clamped values
@@ -87,9 +33,8 @@ func ClampBudget(value int, modelInfo *registry.ModelInfo, provider string) int
 //   - Budget-only model + Level config → Level converted to Budget
 //   - Level-only model + Budget config → Budget converted to Level
 //   - Hybrid model → preserve original format
-func ValidateConfig(config ThinkingConfig, modelInfo *registry.ModelInfo, provider string) (*ThinkingConfig, error) {
+func ValidateConfig(config ThinkingConfig, modelInfo *registry.ModelInfo, fromFormat, toFormat string) (*ThinkingConfig, error) {
-	normalized := config
+	fromFormat, toFormat = strings.ToLower(strings.TrimSpace(fromFormat)), strings.ToLower(strings.TrimSpace(toFormat))
-
 	model := "unknown"
 	support := (*registry.ThinkingSupport)(nil)
 	if modelInfo != nil {
@@ -103,101 +48,108 @@ func ValidateConfig(config ThinkingConfig, modelInfo *registry.ModelInfo, provid
 		if config.Mode != ModeNone {
 			return nil, NewThinkingErrorWithModel(ErrThinkingNotSupported, "thinking not supported for this model", model)
 		}
-		return &normalized, nil
+		return &config, nil
 	}
 
+	allowClampUnsupported := isBudgetBasedProvider(fromFormat) && isLevelBasedProvider(toFormat)
+	strictBudget := fromFormat != "" && isSameProviderFamily(fromFormat, toFormat)
+	budgetDerivedFromLevel := false
+
 	capability := detectModelCapability(modelInfo)
 	switch capability {
 	case CapabilityBudgetOnly:
-		if normalized.Mode == ModeLevel {
+		if config.Mode == ModeLevel {
-			if normalized.Level == LevelAuto {
+			if config.Level == LevelAuto {
 				break
 			}
-			budget, ok := ConvertLevelToBudget(string(normalized.Level))
+			budget, ok := ConvertLevelToBudget(string(config.Level))
 			if !ok {
-				return nil, NewThinkingError(ErrUnknownLevel, fmt.Sprintf("unknown level: %s", normalized.Level))
+				return nil, NewThinkingError(ErrUnknownLevel, fmt.Sprintf("unknown level: %s", config.Level))
 			}
-			normalized.Mode = ModeBudget
+			config.Mode = ModeBudget
-			normalized.Budget = budget
+			config.Budget = budget
-			normalized.Level = ""
+			config.Level = ""
+			budgetDerivedFromLevel = true
 		}
 	case CapabilityLevelOnly:
-		if normalized.Mode == ModeBudget {
+		if config.Mode == ModeBudget {
-			level, ok := ConvertBudgetToLevel(normalized.Budget)
+			level, ok := ConvertBudgetToLevel(config.Budget)
 			if !ok {
-				return nil, NewThinkingError(ErrUnknownLevel, fmt.Sprintf("budget %d cannot be converted to a valid level", normalized.Budget))
+				return nil, NewThinkingError(ErrUnknownLevel, fmt.Sprintf("budget %d cannot be converted to a valid level", config.Budget))
 			}
-			normalized.Mode = ModeLevel
+			// When converting Budget -> Level for level-only models, clamp the derived standard level
-			normalized.Level = ThinkingLevel(level)
+			// to the nearest supported level. Special values (none/auto) are preserved.
-			normalized.Budget = 0
+			config.Mode = ModeLevel
+			config.Level = clampLevel(ThinkingLevel(level), modelInfo, toFormat)
+			config.Budget = 0
 		}
 	case CapabilityHybrid:
 	}
 
-	if normalized.Mode == ModeLevel && normalized.Level == LevelNone {
+	if config.Mode == ModeLevel && config.Level == LevelNone {
-		normalized.Mode = ModeNone
+		config.Mode = ModeNone
-		normalized.Budget = 0
+		config.Budget = 0
-		normalized.Level = ""
+		config.Level = ""
 	}
-	if normalized.Mode == ModeLevel && normalized.Level == LevelAuto {
+	if config.Mode == ModeLevel && config.Level == LevelAuto {
-		normalized.Mode = ModeAuto
+		config.Mode = ModeAuto
-		normalized.Budget = -1
+		config.Budget = -1
-		normalized.Level = ""
+		config.Level = ""
 	}
-	if normalized.Mode == ModeBudget && normalized.Budget == 0 {
+	if config.Mode == ModeBudget && config.Budget == 0 {
-		normalized.Mode = ModeNone
+		config.Mode = ModeNone
-		normalized.Level = ""
+		config.Level = ""
 	}
 
-	if len(support.Levels) > 0 && normalized.Mode == ModeLevel {
+	if len(support.Levels) > 0 && config.Mode == ModeLevel {
-		if !isLevelSupported(string(normalized.Level), support.Levels) {
+		if !isLevelSupported(string(config.Level), support.Levels) {
-			validLevels := normalizeLevels(support.Levels)
+			if allowClampUnsupported {
-			message := fmt.Sprintf("level %q not supported, valid levels: %s", strings.ToLower(string(normalized.Level)), strings.Join(validLevels, ", "))
+				config.Level = clampLevel(config.Level, modelInfo, toFormat)
-			return nil, NewThinkingError(ErrLevelNotSupported, message)
+			}
+			if !isLevelSupported(string(config.Level), support.Levels) {
+				// User explicitly specified an unsupported level - return error
+				// (budget-derived levels may be clamped based on source format)
+				validLevels := normalizeLevels(support.Levels)
+				message := fmt.Sprintf("level %q not supported, valid levels: %s", strings.ToLower(string(config.Level)), strings.Join(validLevels, ", "))
+				return nil, NewThinkingError(ErrLevelNotSupported, message)
+			}
+		}
+	}
+
+	if strictBudget && config.Mode == ModeBudget && !budgetDerivedFromLevel {
+		min, max := support.Min, support.Max
+		if min != 0 || max != 0 {
+			if config.Budget < min || config.Budget > max || (config.Budget == 0 && !support.ZeroAllowed) {
+				message := fmt.Sprintf("budget %d out of range [%d,%d]", config.Budget, min, max)
+				return nil, NewThinkingError(ErrBudgetOutOfRange, message)
+			}
 		}
 	}
 
 	// Convert ModeAuto to mid-range if dynamic not allowed
-	if normalized.Mode == ModeAuto && !support.DynamicAllowed {
+	if config.Mode == ModeAuto && !support.DynamicAllowed {
-		normalized = convertAutoToMidRange(normalized, support, provider, model)
+		config = convertAutoToMidRange(config, support, toFormat, model)
 	}
 
-	if normalized.Mode == ModeNone && provider == "claude" {
+	if config.Mode == ModeNone && toFormat == "claude" {
 		// Claude supports explicit disable via thinking.type="disabled".
 		// Keep Budget=0 so applier can omit budget_tokens.
-		normalized.Budget = 0
+		config.Budget = 0
-		normalized.Level = ""
+		config.Level = ""
 	} else {
-		switch normalized.Mode {
+		switch config.Mode {
 		case ModeBudget, ModeAuto, ModeNone:
-			normalized.Budget = ClampBudget(normalized.Budget, modelInfo, provider)
+			config.Budget = clampBudget(config.Budget, modelInfo, toFormat)
 		}
 
 		// ModeNone with clamped Budget > 0: set Level to lowest for Level-only/Hybrid models
 		// This ensures Apply layer doesn't need to access support.Levels
-		if normalized.Mode == ModeNone && normalized.Budget > 0 && len(support.Levels) > 0 {
+		if config.Mode == ModeNone && config.Budget > 0 && len(support.Levels) > 0 {
-			normalized.Level = ThinkingLevel(support.Levels[0])
+			config.Level = ThinkingLevel(support.Levels[0])
 		}
 	}
 
-	return &normalized, nil
+	return &config, nil
-}
-
-func isLevelSupported(level string, supported []string) bool {
-	for _, candidate := range supported {
-		if strings.EqualFold(level, strings.TrimSpace(candidate)) {
-			return true
-		}
-	}
-	return false
-}
-
-func normalizeLevels(levels []string) []string {
-	normalized := make([]string, 0, len(levels))
-	for _, level := range levels {
-		normalized = append(normalized, strings.ToLower(strings.TrimSpace(level)))
-	}
-	return normalized
 }
 
 // convertAutoToMidRange converts ModeAuto to a mid-range value when dynamic is not allowed.
@@ -246,7 +198,172 @@ func convertAutoToMidRange(config ThinkingConfig, support *registry.ThinkingSupp
 	return config
 }
 
-// logClamp logs a debug message when budget clamping occurs.
+// standardLevelOrder defines the canonical ordering of thinking levels from lowest to highest.
+var standardLevelOrder = []ThinkingLevel{LevelMinimal, LevelLow, LevelMedium, LevelHigh, LevelXHigh}
+
+// clampLevel clamps the given level to the nearest supported level.
+// On tie, prefers the lower level.
+func clampLevel(level ThinkingLevel, modelInfo *registry.ModelInfo, provider string) ThinkingLevel {
+	model := "unknown"
+	var supported []string
+	if modelInfo != nil {
+		if modelInfo.ID != "" {
+			model = modelInfo.ID
+		}
+		if modelInfo.Thinking != nil {
+			supported = modelInfo.Thinking.Levels
+		}
+	}
+
+	if len(supported) == 0 || isLevelSupported(string(level), supported) {
+		return level
+	}
+
+	pos := levelIndex(string(level))
+	if pos == -1 {
+		return level
+	}
+	bestIdx, bestDist := -1, len(standardLevelOrder)+1
+
+	for _, s := range supported {
+		if idx := levelIndex(strings.TrimSpace(s)); idx != -1 {
+			if dist := abs(pos - idx); dist < bestDist || (dist == bestDist && idx < bestIdx) {
+				bestIdx, bestDist = idx, dist
+			}
+		}
+	}
+
+	if bestIdx >= 0 {
+		clamped := standardLevelOrder[bestIdx]
+		log.WithFields(log.Fields{
+			"provider":       provider,
+			"model":          model,
+			"original_level": string(level),
+			"clamped_to":     string(clamped),
+		}).Debug("thinking: level clamped |")
+		return clamped
+	}
+	return level
+}
+
+// clampBudget clamps a budget value to the model's supported range.
+func clampBudget(value int, modelInfo *registry.ModelInfo, provider string) int {
+	model := "unknown"
+	support := (*registry.ThinkingSupport)(nil)
+	if modelInfo != nil {
+		if modelInfo.ID != "" {
+			model = modelInfo.ID
+		}
+		support = modelInfo.Thinking
+	}
+	if support == nil {
+		return value
+	}
+
+	// Auto value (-1) passes through without clamping.
+	if value == -1 {
+		return value
+	}
+
+	min, max := support.Min, support.Max
+	if value == 0 && !support.ZeroAllowed {
+		log.WithFields(log.Fields{
+			"provider":       provider,
+			"model":          model,
+			"original_value": value,
+			"clamped_to":     min,
+			"min":            min,
+			"max":            max,
+		}).Warn("thinking: budget zero not allowed |")
+		return min
+	}
+
+	// Some models are level-only and do not define numeric budget ranges.
+	if min == 0 && max == 0 {
+		return value
+	}
+
+	if value < min {
+		if value == 0 && support.ZeroAllowed {
+			return 0
+		}
+		logClamp(provider, model, value, min, min, max)
+		return min
+	}
+	if value > max {
+		logClamp(provider, model, value, max, min, max)
+		return max
+	}
+	return value
+}
+
+func isLevelSupported(level string, supported []string) bool {
+	for _, s := range supported {
+		if strings.EqualFold(level, strings.TrimSpace(s)) {
+			return true
+		}
+	}
+	return false
+}
+
+func levelIndex(level string) int {
+	for i, l := range standardLevelOrder {
+		if strings.EqualFold(level, string(l)) {
+			return i
+		}
+	}
+	return -1
+}
+
+func normalizeLevels(levels []string) []string {
+	out := make([]string, len(levels))
+	for i, l := range levels {
+		out[i] = strings.ToLower(strings.TrimSpace(l))
+	}
+	return out
+}
+
+func isBudgetBasedProvider(provider string) bool {
+	switch provider {
+	case "gemini", "gemini-cli", "antigravity", "claude":
+		return true
+	default:
+		return false
+	}
+}
+
+func isLevelBasedProvider(provider string) bool {
+	switch provider {
+	case "openai", "openai-response", "codex":
+		return true
+	default:
+		return false
+	}
+}
+
+func isGeminiFamily(provider string) bool {
+	switch provider {
+	case "gemini", "gemini-cli", "antigravity":
+		return true
+	default:
+		return false
+	}
+}
+
+func isSameProviderFamily(from, to string) bool {
+	if from == to {
+		return true
+	}
+	return isGeminiFamily(from) && isGeminiFamily(to)
+}
+
+func abs(x int) int {
+	if x < 0 {
+		return -x
+	}
+	return x
+}
+
 func logClamp(provider, model string, original, clampedTo, min, max int) {
 	log.WithFields(log.Fields{
 		"provider":       provider,

internal/translator/antigravity/claude/antigravity_claude_request.go

@@ -12,7 +12,6 @@ import (
 	"strings"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/cache"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/translator/gemini/common"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
@@ -388,14 +387,11 @@ func ConvertClaudeRequestToAntigravity(modelName string, inputRawJSON []byte, _
 
 	// Map Anthropic thinking -> Gemini thinkingBudget/include_thoughts when type==enabled
 	if t := gjson.GetBytes(rawJSON, "thinking"); t.Exists() && t.IsObject() {
-		modelInfo := registry.LookupModelInfo(modelName)
+		if t.Get("type").String() == "enabled" {
-		if modelInfo != nil && modelInfo.Thinking != nil {
+			if b := t.Get("budget_tokens"); b.Exists() && b.Type == gjson.Number {
-			if t.Get("type").String() == "enabled" {
+				budget := int(b.Int())
-				if b := t.Get("budget_tokens"); b.Exists() && b.Type == gjson.Number {
+				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
-					budget := int(b.Int())
+				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
-					out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
-					out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.include_thoughts", true)
-				}
 			}
 		}
 	}

internal/translator/antigravity/claude/antigravity_claude_request_test.go

@@ -343,8 +343,8 @@ func TestConvertClaudeRequestToAntigravity_ThinkingConfig(t *testing.T) {
 		if thinkingConfig.Get("thinkingBudget").Int() != 8000 {
 			t.Errorf("Expected thinkingBudget 8000, got %d", thinkingConfig.Get("thinkingBudget").Int())
 		}
-		if !thinkingConfig.Get("include_thoughts").Bool() {
+		if !thinkingConfig.Get("includeThoughts").Bool() {
-			t.Error("include_thoughts should be true")
+			t.Error("includeThoughts should be true")
 		}
 	} else {
 		t.Log("thinkingConfig not present - model may not be registered in test registry")

internal/translator/claude/gemini/claude_gemini_request.go

@@ -15,7 +15,7 @@ import (
 	"strings"
 
 	"github.com/google/uuid"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
+	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
@@ -115,18 +115,41 @@ func ConvertGeminiRequestToClaude(modelName string, inputRawJSON []byte, stream
 			}
 		}
 		// Include thoughts configuration for reasoning process visibility
-		// Only apply for models that support thinking and use numeric budgets, not discrete levels.
+		// Translator only does format conversion, ApplyThinking handles model capability validation.
 		if thinkingConfig := genConfig.Get("thinkingConfig"); thinkingConfig.Exists() && thinkingConfig.IsObject() {
-			modelInfo := registry.LookupModelInfo(modelName)
+			if thinkingLevel := thinkingConfig.Get("thinkingLevel"); thinkingLevel.Exists() {
-			if modelInfo != nil && modelInfo.Thinking != nil && len(modelInfo.Thinking.Levels) == 0 {
+				level := strings.ToLower(strings.TrimSpace(thinkingLevel.String()))
-				// Check for thinkingBudget first - if present, enable thinking with budget
+				switch level {
-				if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() && thinkingBudget.Int() > 0 {
+				case "":
-					out, _ = sjson.Set(out, "thinking.type", "enabled")
+				case "none":
-					out, _ = sjson.Set(out, "thinking.budget_tokens", thinkingBudget.Int())
+					out, _ = sjson.Set(out, "thinking.type", "disabled")
-				} else if includeThoughts := thinkingConfig.Get("include_thoughts"); includeThoughts.Exists() && includeThoughts.Type == gjson.True {
+					out, _ = sjson.Delete(out, "thinking.budget_tokens")
-					// Fallback to include_thoughts if no budget specified
+				case "auto":
 					out, _ = sjson.Set(out, "thinking.type", "enabled")
+					out, _ = sjson.Delete(out, "thinking.budget_tokens")
+				default:
+					if budget, ok := thinking.ConvertLevelToBudget(level); ok {
+						out, _ = sjson.Set(out, "thinking.type", "enabled")
+						out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
+					}
 				}
+			} else if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() {
+				budget := int(thinkingBudget.Int())
+				switch budget {
+				case 0:
+					out, _ = sjson.Set(out, "thinking.type", "disabled")
+					out, _ = sjson.Delete(out, "thinking.budget_tokens")
+				case -1:
+					out, _ = sjson.Set(out, "thinking.type", "enabled")
+					out, _ = sjson.Delete(out, "thinking.budget_tokens")
+				default:
+					out, _ = sjson.Set(out, "thinking.type", "enabled")
+					out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
+				}
+			} else if includeThoughts := thinkingConfig.Get("includeThoughts"); includeThoughts.Exists() && includeThoughts.Type == gjson.True {
+				out, _ = sjson.Set(out, "thinking.type", "enabled")
+			} else if includeThoughts := thinkingConfig.Get("include_thoughts"); includeThoughts.Exists() && includeThoughts.Type == gjson.True {
+				out, _ = sjson.Set(out, "thinking.type", "enabled")
 			}
 		}
 	}

internal/translator/claude/openai/chat-completions/claude_openai_request.go

@@ -15,7 +15,6 @@ import (
 	"strings"
 
 	"github.com/google/uuid"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
@@ -66,23 +65,21 @@ func ConvertOpenAIRequestToClaude(modelName string, inputRawJSON []byte, stream
 
 	root := gjson.ParseBytes(rawJSON)
 
+	// Convert OpenAI reasoning_effort to Claude thinking config.
 	if v := root.Get("reasoning_effort"); v.Exists() {
-		modelInfo := registry.LookupModelInfo(modelName)
+		effort := strings.ToLower(strings.TrimSpace(v.String()))
-		if modelInfo != nil && modelInfo.Thinking != nil && len(modelInfo.Thinking.Levels) == 0 {
+		if effort != "" {
-			effort := strings.ToLower(strings.TrimSpace(v.String()))
+			budget, ok := thinking.ConvertLevelToBudget(effort)
-			if effort != "" {
+			if ok {
-				budget, ok := thinking.ConvertLevelToBudget(effort)
+				switch budget {
-				if ok {
+				case 0:
-					switch budget {
+					out, _ = sjson.Set(out, "thinking.type", "disabled")
-					case 0:
+				case -1:
-						out, _ = sjson.Set(out, "thinking.type", "disabled")
+					out, _ = sjson.Set(out, "thinking.type", "enabled")
-					case -1:
+				default:
+					if budget > 0 {
 						out, _ = sjson.Set(out, "thinking.type", "enabled")
-					default:
+						out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
-						if budget > 0 {
-							out, _ = sjson.Set(out, "thinking.type", "enabled")
-							out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
-						}
 					}
 				}
 			}

internal/translator/claude/openai/responses/claude_openai-responses_request.go

@@ -10,7 +10,6 @@ import (
 	"strings"
 
 	"github.com/google/uuid"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
@@ -54,23 +53,21 @@ func ConvertOpenAIResponsesRequestToClaude(modelName string, inputRawJSON []byte
 
 	root := gjson.ParseBytes(rawJSON)
 
+	// Convert OpenAI Responses reasoning.effort to Claude thinking config.
 	if v := root.Get("reasoning.effort"); v.Exists() {
-		modelInfo := registry.LookupModelInfo(modelName)
+		effort := strings.ToLower(strings.TrimSpace(v.String()))
-		if modelInfo != nil && modelInfo.Thinking != nil && len(modelInfo.Thinking.Levels) == 0 {
+		if effort != "" {
-			effort := strings.ToLower(strings.TrimSpace(v.String()))
+			budget, ok := thinking.ConvertLevelToBudget(effort)
-			if effort != "" {
+			if ok {
-				budget, ok := thinking.ConvertLevelToBudget(effort)
+				switch budget {
-				if ok {
+				case 0:
-					switch budget {
+					out, _ = sjson.Set(out, "thinking.type", "disabled")
-					case 0:
+				case -1:
-						out, _ = sjson.Set(out, "thinking.type", "disabled")
+					out, _ = sjson.Set(out, "thinking.type", "enabled")
-					case -1:
+				default:
+					if budget > 0 {
 						out, _ = sjson.Set(out, "thinking.type", "enabled")
-					default:
+						out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
-						if budget > 0 {
-							out, _ = sjson.Set(out, "thinking.type", "enabled")
-							out, _ = sjson.Set(out, "thinking.budget_tokens", budget)
-						}
 					}
 				}
 			}

internal/translator/codex/claude/codex_claude_request.go

@@ -12,7 +12,6 @@ import (
 	"strings"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/misc"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
@@ -218,18 +217,15 @@ func ConvertClaudeRequestToCodex(modelName string, inputRawJSON []byte, _ bool)
 	// Add additional configuration parameters for the Codex API.
 	template, _ = sjson.Set(template, "parallel_tool_calls", true)
 
-	// Convert thinking.budget_tokens to reasoning.effort for level-based models
+	// Convert thinking.budget_tokens to reasoning.effort.
-	reasoningEffort := "medium" // default
+	reasoningEffort := "medium"
 	if thinkingConfig := rootResult.Get("thinking"); thinkingConfig.Exists() && thinkingConfig.IsObject() {
-		modelInfo := registry.LookupModelInfo(modelName)
 		switch thinkingConfig.Get("type").String() {
 		case "enabled":
-			if modelInfo != nil && modelInfo.Thinking != nil && len(modelInfo.Thinking.Levels) > 0 {
+			if budgetTokens := thinkingConfig.Get("budget_tokens"); budgetTokens.Exists() {
-				if budgetTokens := thinkingConfig.Get("budget_tokens"); budgetTokens.Exists() {
+				budget := int(budgetTokens.Int())
-					budget := int(budgetTokens.Int())
+				if effort, ok := thinking.ConvertBudgetToLevel(budget); ok && effort != "" {
-					if effort, ok := thinking.ConvertBudgetToLevel(budget); ok && effort != "" {
+					reasoningEffort = effort
-						reasoningEffort = effort
-					}
 				}
 			}
 		case "disabled":

internal/translator/codex/gemini/codex_gemini_request.go

@@ -14,7 +14,6 @@ import (
 	"strings"
 
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/misc"
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/thinking"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
 	"github.com/tidwall/gjson"
@@ -249,22 +248,28 @@ func ConvertGeminiRequestToCodex(modelName string, inputRawJSON []byte, _ bool)
 	// Fixed flags aligning with Codex expectations
 	out, _ = sjson.Set(out, "parallel_tool_calls", true)
 
-	// Convert thinkingBudget to reasoning.effort for level-based models
+	// Convert Gemini thinkingConfig to Codex reasoning.effort.
-	reasoningEffort := "medium" // default
+	effortSet := false
 	if genConfig := root.Get("generationConfig"); genConfig.Exists() {
 		if thinkingConfig := genConfig.Get("thinkingConfig"); thinkingConfig.Exists() && thinkingConfig.IsObject() {
-			modelInfo := registry.LookupModelInfo(modelName)
+			if thinkingLevel := thinkingConfig.Get("thinkingLevel"); thinkingLevel.Exists() {
-			if modelInfo != nil && modelInfo.Thinking != nil && len(modelInfo.Thinking.Levels) > 0 {
+				effort := strings.ToLower(strings.TrimSpace(thinkingLevel.String()))
-				if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() {
+				if effort != "" {
-					budget := int(thinkingBudget.Int())
+					out, _ = sjson.Set(out, "reasoning.effort", effort)
-					if effort, ok := thinking.ConvertBudgetToLevel(budget); ok && effort != "" {
+					effortSet = true
-						reasoningEffort = effort
+				}
-					}
+			} else if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() {
+				if effort, ok := thinking.ConvertBudgetToLevel(int(thinkingBudget.Int())); ok {
+					out, _ = sjson.Set(out, "reasoning.effort", effort)
+					effortSet = true
 				}
 			}
 		}
 	}
-	out, _ = sjson.Set(out, "reasoning.effort", reasoningEffort)
+	if !effortSet {
+		// No thinking config, set default effort
+		out, _ = sjson.Set(out, "reasoning.effort", "medium")
+	}
 	out, _ = sjson.Set(out, "reasoning.summary", "auto")
 	out, _ = sjson.Set(out, "stream", true)
 	out, _ = sjson.Set(out, "store", false)

internal/translator/gemini-cli/claude/gemini-cli_claude_request.go

@@ -9,7 +9,6 @@ import (
 	"bytes"
 	"strings"
 
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/translator/gemini/common"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
@@ -161,14 +160,11 @@ func ConvertClaudeRequestToCLI(modelName string, inputRawJSON []byte, _ bool) []
 
 	// Map Anthropic thinking -> Gemini thinkingBudget/include_thoughts when type==enabled
 	if t := gjson.GetBytes(rawJSON, "thinking"); t.Exists() && t.IsObject() {
-		modelInfo := registry.LookupModelInfo(modelName)
+		if t.Get("type").String() == "enabled" {
-		if modelInfo != nil && modelInfo.Thinking != nil {
+			if b := t.Get("budget_tokens"); b.Exists() && b.Type == gjson.Number {
-			if t.Get("type").String() == "enabled" {
+				budget := int(b.Int())
-				if b := t.Get("budget_tokens"); b.Exists() && b.Type == gjson.Number {
+				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
-					budget := int(b.Int())
+				out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.includeThoughts", true)
-					out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.thinkingBudget", budget)
-					out, _ = sjson.Set(out, "request.generationConfig.thinkingConfig.include_thoughts", true)
-				}
 			}
 		}
 	}

internal/translator/gemini/claude/gemini_claude_request.go

@@ -9,7 +9,6 @@ import (
 	"bytes"
 	"strings"
 
-	"github.com/router-for-me/CLIProxyAPI/v6/internal/registry"
 	"github.com/router-for-me/CLIProxyAPI/v6/internal/translator/gemini/common"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
@@ -153,16 +152,13 @@ func ConvertClaudeRequestToGemini(modelName string, inputRawJSON []byte, _ bool)
 	}
 
 	// Map Anthropic thinking -> Gemini thinkingBudget/include_thoughts when enabled
-	// Only apply for models that use numeric budgets, not discrete levels.
+	// Translator only does format conversion, ApplyThinking handles model capability validation.
 	if t := gjson.GetBytes(rawJSON, "thinking"); t.Exists() && t.IsObject() {
-		modelInfo := registry.LookupModelInfo(modelName)
+		if t.Get("type").String() == "enabled" {
-		if modelInfo != nil && modelInfo.Thinking != nil && len(modelInfo.Thinking.Levels) == 0 {
+			if b := t.Get("budget_tokens"); b.Exists() && b.Type == gjson.Number {
-			if t.Get("type").String() == "enabled" {
+				budget := int(b.Int())
-				if b := t.Get("budget_tokens"); b.Exists() && b.Type == gjson.Number {
+				out, _ = sjson.Set(out, "generationConfig.thinkingConfig.thinkingBudget", budget)
-					budget := int(b.Int())
+				out, _ = sjson.Set(out, "generationConfig.thinkingConfig.includeThoughts", true)
-					out, _ = sjson.Set(out, "generationConfig.thinkingConfig.thinkingBudget", budget)
-					out, _ = sjson.Set(out, "generationConfig.thinkingConfig.include_thoughts", true)
-				}
 			}
 		}
 	}

internal/translator/openai/gemini/openai_gemini_request.go

@@ -77,12 +77,15 @@ func ConvertGeminiRequestToOpenAI(modelName string, inputRawJSON []byte, stream
 			}
 		}
 
-		// Convert thinkingBudget to reasoning_effort
+		// Map Gemini thinkingConfig to OpenAI reasoning_effort.
-		// Always perform conversion to support allowCompat models that may not be in registry
 		if thinkingConfig := genConfig.Get("thinkingConfig"); thinkingConfig.Exists() && thinkingConfig.IsObject() {
-			if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() {
+			if thinkingLevel := thinkingConfig.Get("thinkingLevel"); thinkingLevel.Exists() {
-				budget := int(thinkingBudget.Int())
+				effort := strings.ToLower(strings.TrimSpace(thinkingLevel.String()))
-				if effort, ok := thinking.ConvertBudgetToLevel(budget); ok && effort != "" {
+				if effort != "" {
+					out, _ = sjson.Set(out, "reasoning_effort", effort)
+				}
+			} else if thinkingBudget := thinkingConfig.Get("thinkingBudget"); thinkingBudget.Exists() {
+				if effort, ok := thinking.ConvertBudgetToLevel(int(thinkingBudget.Int())); ok {
 					out, _ = sjson.Set(out, "reasoning_effort", effort)
 				}
 			}